AI coding stack emerges

Developers are starting to split coding work across three assistants instead of betting on one. Cursor is becoming the control panel, Claude Code the planner and reviewer, and OpenAI Codex the executor. (thenewstack.io) That pattern sharpened in early April 2026 as Cursor shipped Cursor 3 with an “Agents Window” for running many agents in parallel across local machines, cloud environments, remote Secure Shell sessions, and multiple repositories. Cursor said the release became generally available on April 2, 2026. (cursor.com) Anthropic positions Claude Code as a system that can read a codebase, change files, run tests, and deliver committed code, and it added enterprise controls in August 2025 so teams could manage usage under business plans. The company also launched a multi-agent code review feature for Team and Enterprise customers last month. (anthropic.com 1) (anthropic.com 2) (thenewstack.io) OpenAI now describes Codex as its coding agent for software development and says it is included with ChatGPT Plus, Pro, Business, Education, and Enterprise plans. OpenAI’s developer docs also describe subagents that can be spawned on request and coordinated into a single result. (developers.openai.com 1) (developers.openai.com 2) The plumbing behind these handoffs is a standard called Model Context Protocol, which works like a universal adapter for assistants and external tools. Anthropic introduced it in November 2024, and in December 2025 said it was donating the protocol to the Agentic Artificial Intelligence Foundation under the Linux Foundation with OpenAI, Block, and others backing the effort. (anthropic.com 1) (anthropic.com 2) That standard is now showing up in the small utilities developers are building around the big coding tools. A developer writing under the name imviky-ctrl published a Model Context Protocol server on April 12, 2026 that tells Claude when ChatGPT is down, what different models cost, and which application programming interfaces are available. (dev.to) The point of that server is not code generation itself. It gives an assistant operational context — status pages, pricing tables, and service metadata — so the model can decide whether to call one tool, switch to another, or warn the user about cost before it starts. (dev.to) (anthropic.com) The ecosystem around that adapter layer is expanding fast. The official Model Context Protocol registry shows newly updated servers every week, and Anthropic said in January that the protocol had reached 100 million monthly downloads and become its standard way to connect models to tools and data. (registry.modelcontextprotocol.io) (anthropic.com) Vendors are also moving toward the same shape from different directions. Cursor is centering its product on agent management, Anthropic is adding more autonomous coding and review features, and OpenAI is adding plugins and editor integrations so Codex can work across command line tools, apps, and code editors. (cursor.com) (anthropic.com) (thenewstack.io) (developers.openai.com) What is emerging looks less like one winner-take-all coding assistant and more like a stack: one tool to coordinate work, one to reason through it, one to execute it, and a growing layer of software that tells each assistant what is available, what is broken, and what it will cost. (thenewstack.io) (dev.to)